You have a harness you never meant to build

A creator ran an inventory on his own AI setup this week and did not like what he found. One ordinary writing job was dragging in an 18,000-word file before it even saw his prompt. He counted 66 skills and 172 instruction files, and one rule he cared about, keeping the AI from putting words in his mouth, existed in 15 different versions that had quietly drifted out of sync.

None of it was added carelessly. That is the trap. Every rule fixed a real problem the day it was written. The model missed something, so he added an instruction. It missed something else, so he added another. Each was reasonable, and the sum was a harness so heavy it caused the next miss, which earned another rule.

A harness, plainly, is everything wrapped around the model: the instructions, the memory, the saved prompts, the tools, the permissions. It shapes the answer before you type a word. And it grows by accretion, one correction at a time, until nobody can see the whole thing at once.

The vault’s own rule for this is context budget discipline: do not stuff every note into the window, route to the smallest reliable slice and load the specialist material only when the work reaches that phase. He found the same thing the hard way. His thick setup produced richer analysis and then failed the actual delivery twice, once breaking the format, once the length. The lean one passed every time.

Having said that, the fix is not shorter is better, and he is careful to say so. Some of that bloat was load-bearing. The fix is being able to see the system, so you can tell which rules protect the work and which are just barnacles. The models keep changing underneath you. The correction you gave six months ago is still in there, quietly shaping answers for a problem that no longer exists.